Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Being Bayesian about Categorical Probability
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace softmax cross-entropy with belief matching | ResNet-18 and ResNet-50 · CIFAR-10 and CIFAR-100 | classification error: 6.13 ± 0.13 -> 5.93 ± 0.07; 26.44 ± 0.33 -> 24.19 ± 0.34; 5.76 ± 0.06 -> 5.59 ± 0.05; 25.00 ± 0.23 -> 23.86 ± 0.37 | Helped |
| Replace softmax cross-entropy with belief matching | ResNext-50 and ResNext-101 · ImageNet validation | Top1 and Top5 error: 22.23 -> 22.03; 6.36 -> 6.29; 20.72 -> 20.23; 5.59 -> 5.26 | Helped |
| Replace softmax cross-entropy with belief matching | ImageNet-pretrained ResNet-50 fine-tuned on downstream tasks · CIFAR-10, Food-101, Cars | classification error: 5.44 ± 0.10 -> 5.03 ± 0.04; 28.49 ± 0.08 -> 26.41 ± 0.07; 42.99 ± 0.14 -> 39.99 ± 0.20 | Helped |
| Use Dirichlet distribution matching instead of categorical-probability matching for consistency | Wide ResNet 28-2 with Pi-Model and VAT · CIFAR-10 semi-supervised | classification error: 16.52 ± 0.21 -> 16.01 ± 0.36; 13.33 ± 0.37 -> 12.40 ± 0.23 | Helped |
| Add the KL prior-regularization term (ablation) | BM with ResNet-50 · CIFAR-10 | classification error: 5.68% without the KL term; full-method value not given in this comparison | Helped |
| Add the KL prior-regularization term (ablation) | BM with ResNet-50 · CIFAR-100 | classification error: 24.69% without the KL term; full-method value not given in this comparison | Helped |
| Remove the prior KL term while retaining Dirichlet distribution matching (ablation) | BM with ResNet-50 · CIFAR-10 and CIFAR-100 | classification error: 5.68% on CIFAR-10; 24.69% on CIFAR-100 | Mixed |
| Use BM rather than softmax cross-entropy | ResNet-50 · CIFAR-10 and CIFAR-100 calibration | ECE: 3.82 -> 1.66; 13.48 -> 4.25 | Helped |
| Use BM instead of softmax | ResNet-50 · OOD samples, including SVHN | predictive uncertainty | Helped |
| Use BM instead of softmax | ImageNet-pretrained ResNet-50 fine-tuned only in its last layer · OOD transfer-learning datasets | predictive uncertainty | Helped |
| Use MC Drop (last) instead of softmax | ResNet-18 · CIFAR-10 test | classification error: 6.13 ± 0.13 -> 6.13 ± 0.08 | No clear effect |
| Apply MC Dropout to all layers instead of using softmax | ResNet-18 · CIFAR-10 and CIFAR-100 test | classification error: 6.13 ± 0.13 -> 6.50 ± 0.14; 26.44 ± 0.33 -> 27.32 ± 0.45 | Hurt |
Closest papers by meaning
- One-vs-Each Approximation to Softmax for Scalable Estimation of Probabilities (NeurIPS 2016 · similarity 0.87)
- Bayesian Deep Learning and a Probabilistic Perspective of Generalization (NeurIPS 2020 · similarity 0.87)
- Beyond Softmax: A Natural Parameterization for Categorical Random Variables (ICML 2026 · similarity 0.87)
- Density-Softmax: Efficient Test-time Model for Uncertainty Estimation and Robustness under Distribution Shifts (ICML 2024 · similarity 0.87)
- Rethinking Approximate Gaussian Inference in Classification (NeurIPS 2025 · similarity 0.86)
- Dirichlet-Based Prediction Calibration for Learning with Noisy Labels (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.